Skip to content

SYSTEM Cited by 1 source

Flattened Planar Fabric (FPF)

Flattened Planar Fabric (FPF) is Meta's multiplane backend scale-out fabric — the network that connects GPUs to one another across a training cluster — introduced at OCP Summit 2026. It "fundamentally changes how GPUs connect to the network and the network control plane of the fabric," replacing the traditional single-plane uplink architecture of Meta's prior NSF fabric.

The core idea: one GPU, eight planes

With FPF, each GPU connects via eight independent 100G links to eight different switches — one per network plane. Traffic from a GPU is spread across all eight planes rather than funneled through a single uplink.

Four advantages Meta claims

  1. Larger non-blocking GPU clusters. The multiplane layout supports bigger non-blocking clusters than single-plane topologies — the scale Meta's largest training workloads need.
  2. A flattened topology — 3-tier → 2-tier Clos. FPF moves from the 3-tier Clos of the NSF fabric to a 2-tier Clos. Removing a switch tier cuts network hops, latency, and infrastructure cost.
  3. NIC-level hardware load balancing. Load balancing — traditionally done by network switches (ECMP) — now starts at the NIC, which sprays every packet across all eight planes in hardware. (Contrast switch-ECMP's fat-flow problem: a few long-lived tensor flows hash to one path.)
  4. A combined control plane for fault tolerance. FPF's control plane spans both the switches and the hosts, so a host can steer traffic away from a downed plane — fault handling becomes an endpoint-aware decision rather than a purely in-fabric one.

Built with NVIDIA Spectrum-X

Meta developed FPF running inside its own infrastructure in close collaboration with NVIDIA, leveraging NVIDIA Spectrum-X Multiplane innovations so FPF performs well under assorted failure conditions and operational-lifecycle scenarios. Meta and NVIDIA have also worked with joint cloud partners to help standardize multiplane-fabric implementations based on FPF learnings — part of Meta's effort to make leased/public-cloud GPU capacity behave like its own racks.

Where it sits

FPF is the scale-out layer of Meta's scale-up / scale-out / scale-across decomposition (see concepts/scale-up-scale-out-scale-across). It rides the 102.4T IceCube switch generation, carries lossless RoCE traffic, and is one of the building-scale fabrics the BAG super-spine knits together into a region-scale cluster. It is the planar-multiplane successor to the single-plane NSF and a sibling evolution to DSF.

Relationship to MetaRoCE

FPF's NIC-level per-packet spraying across planes and host-aware plane failover are the fabric-side complement to MetaRoCE's endpoint-side design (per-packet spraying across logical paths, NIC-chosen plane selection, treating a down plane as a reroute rather than a stall). FPF puts the planes in the fabric; MetaRoCE is the transport that drives them from the endpoint.

Caveats

Introduced in an OCP summit announcement — the control-plane protocol between hosts and switches, the exact plane-failover signaling, and non-blocking cluster sizes are described conceptually, not specified.

Seen in (wiki)

Last updated · 854 distilled / 2,448 read